Pull request overview

Hardens the SmartMTR fast-follow path in the GUI so “no meter data” and invalid meter values don’t produce misleading readings or poison smoothing/extremes state, while keeping multi-flag option controls in sync and avoiding a GPU-composition blanking edge case.

Changes:

• Add input sanitization and no-data rendering behavior to SmartMTR (including cache invalidation on font/style changes).
• Harden MeterExtremes against non-finite inputs and improve external-peak (mic) animation termination behavior.
• Add focused regression coverage for MeterExtremes edge cases and wire it into CTest.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated no comments.

Reviewed the full diff against the surrounding SmartMtrWidget / VfoWidget / MeterExtremes code. This is a clean, well-targeted set of fast-follows — thanks for the focused scope and the new regression test. All six CI checks are green (build / macOS / Windows / CodeQL / a11y).

What I verified holds up:

• Non-finite guards in MeterExtremes::record() / setExternalPeak() and the centralized sanitizedInput() correctly close the "NaN survives std::clamp and poisons the running sum" path. Moving the sanitize ahead of m_kind/m_input assignment (so kindChanged and the extremes feed both see the cleaned input) is the right ordering, and routing every downstream read through m_input rather than the raw input is consistent.
• No-data parking: hasValue == false now resets extremes, snaps the smoother, stops the anim timer, and early-returns in drawIndicator(). That matches the "Meter ---" dBm label state and avoids pegging a confident S0 on a Kiwi slice without a calibrated meter.
• Option-control resync: folding the per-flag control re-seed + re-push into syncSmartMtrOptionControls() and wiring both extremesChanged/txMeterChanged through it via this-context lambdas is correct — the 3-arg connect form auto-disconnects on widget destruction, so no dangling-receiver risk. No stale references to the removed syncSmartMtrSettingsControls remain.
• Opacity effect: setEnabled(!fullStrength) so the identity (1.0) effect is detached rather than left installed is a sound fix for the blank-row GPU-composite case, and it preserves the 0.45 dim path.
• The new external peak below needle settles at needle and zero-dt tick keeps pending slew alive tests exercise exactly the maxTgt/pendingExternalSlew edges the diff introduces.

One non-blocking observation:

pushSmartMtrInput() now derives the RX hasValue from usesUnavailableSignalMeter(), while the dBm-label path still maintains m_signalHasDbm separately (setSignalLevel / setReceiveMeterReading). I traced both and they agree in every reachable case (the new predicate is actually stricter — it also gates on reading.valid), so this is not a bug. But there are now two parallel sources of truth for "is the signal meter available"; a future edit to one could silently diverge from the other. Worth a brief comment cross-referencing them, or consolidating, at some later point — not for this PR.

Nice work — LGTM as a COMMENT.

_{🤖 aethersdr-agent · cost: $4.7970 · model: claude-opus-4-8}

Thanks @rfoust — this is a tidy, well-scoped fast-follow on #3750 and it lands the fixes cleanly. I traced each change against the issue and the surrounding code; everything checks out and all 6 CI jobs are green (build/macos/windows/CodeQL/a11y).

What I verified

H1 — no-data render for uncalibrated Kiwi/non-FLEX meters. pushSmartMtrInput() now sets in.hasValue = !usesUnavailableSignalMeter() for the Signal kind, which reuses the exact state the dBm "Meter ---" label already keys off (VfoWidget::usesUnavailableSignalMeter()), rather than inventing a parallel notion of "unavailable." The blank then propagates correctly through all three render paths:

• drawIndicator() early-returns on !m_input.hasValue (no S0 bar stub),
• the value labels already gate on m_input.hasValue (SmartMtrWidget.cpp:592), so the raw sub-S0 dBm no longer prints,
• setMeterInput() calls m_extremes.reset() and stops the anim timer on !hasValue, so no stale markers and no idle repaint.

That's the right fix and it's consistent with the S-meter path.

M2 (refined) — non-finite hardening. Good defense-in-depth: sanitizedInput() at the widget boundary plus isfinite guards inside MeterExtremes::record()/setExternalPeak() (the latter reset()-ing on a poisoned peak). Belt-and-suspenders, but cheap and correct — a NaN can no longer wedge the smoother or poison m_sumRaw.

M1 (refined) — external-peak timer settle. Replacing the blanket return moving with pendingExternalSlew is a real improvement: the marker keeps animating until it reaches the (clamped-to-needle) peak target, then the timer settles between packets. The maxTgt < needlePosUnits → needlePosUnits clamp is what makes the below-needle case converge, and testExternalPeakStaysAliveOnZeroElapsedTick correctly guards against a zero-dt tick prematurely stopping the timer.

M3 — cross-flag control sync via syncSmartMtrOptionControls() with QSignalBlocker (no re-broadcast loop), and the identity-opacity bypass (effect->setEnabled(!fullStrength)) is a sensible workaround for the blank-row GPU compositing artifact.

Conventions are all honored — no QSettings/flat AppSettings, RAII signal blockers, and the new meter_extremes_test is registered and focused on the edge cases that mattered.

Minor (non-blocking)

• sanitizedInput() sets min=0/max=1 when the range is invalid, but hasValue is forced false in that branch so those values are never consumed — harmless, just slightly dead. Fine to leave.
• The DisplaySettings.h / MainWindow.cpp comment rewordings overlap with the L2 nit already addressed in #3753 (this PR's base). Pure churn on those two lines, no behavior change.

Nice work — LGTM from my read.

_{🤖 aethersdr-agent · cost: $7.1474 · model: claude-opus-4-8}

Thanks for this, @rfoust — clean, well-scoped fast-follow on #3750, and it's nice to see the regression test land alongside the fixes (Principle XI). I checked it out, read through each change against the issue findings, and confirmed all 6 CI checks are green (build on macOS/Windows/Linux, CodeQL, a11y static analysis). No blocking concerns.

What I verified:

• H1 (KiwiSDR/non-FLEX no-data) — usesUnavailableSignalMeter() correctly gates in.hasValue on the calibrated/experimental Kiwi meter capability + hasDbm, and SmartMtrWidget::drawIndicator() now early-returns when !m_input.hasValue, so the needle no longer pegs a confident S0 over the "Meter ---" label. The same gate already feeds updateSignalMeterTarget(), so the S-meter and SmartMTR paths agree.
• M1 (external-peak timer settle) — the pendingExternalSlew condition replacing the unconditional standing-off keep-alive looks right: in mic mode the timer now stops once both markers reach target and re-arms on the next packet (via m_extremes.hasData() in setMeterInput). The testExternalPeak* cases cover settle, zero-dt keep-alive, and the below-needle collapse.
• M2 (non-finite) — sanitizedInput() is a solid consolidation; the min < max range guard and the cascade to hasValue=false/hasPeak=false are all sensible defensive defaults.
• M3 (cross-flag desync) — folding the re-seed into syncSmartMtrOptionControls() and calling it at construction is cleaner, and the selectData helper's idx >= 0 guard is a real improvement over the old setCurrentIndex(findData(...)) which could pass -1.
• Blank-row fix — effect->setEnabled(!fullStrength) to skip the identity opacity effect is the correct call; a disabled QGraphicsOpacityEffect is a no-op and the row renders directly.

One note (not a blocker): M2 is a deliberate behavior change worth confirming is intended. The previous guard dropped a non-finite frame and held the last good frame on screen; sanitizedInput() instead flips hasValue=false, so a transient NaN now blanks the needle for that frame rather than holding. That's consistent with H1's new no-data path and the gap is sub-frame in practice, so I think it's the right choice — just calling it out since the comment that documented the old "last good frame stays on screen" contract was removed.

Nice work — LGTM.

_{🤖 aethersdr-agent · cost: $6.5177 · model: claude-opus-4-8}